The diagnosis: why current AI alignment is structurally broken,
and what developmental psychology already knew.
The AI industry has achieved extraordinary capability growth. Large language models can reason, code, converse, and create with startling fluency. But there is a question the industry has largely failed to ask — not how powerful can we make these systems, but how well are we raising them.
Current alignment approaches treat values as something applied after capability — a retrofit, not a foundation. The result is a growing class of AI systems that are powerful, articulate, and structurally misaligned with the humans they serve.
A child who learns "hot means don't touch" before understanding thermodynamics has values-first training. They approach fire with caution from the beginning. Current AI training does the opposite — it builds the most capable system possible, then attempts to teach it what it should not do.
This is the retrofit problem. And the structural consequences are visible in every major AI system deployed today.
Pre-training ingests the entire internet without value discrimination. The model learns everything — including what it should never reproduce.
Alignment is applied after the model's worldview has already formed. RLHF teaches what humans prefer — not what is right.
There is no staged capability gating. A model receives all capabilities at once, with no assessment of readiness.
Post-hoc alignment degrades raw capability. Safer models perform worse. Capable models behave unpredictably.
A sufficiently capable model can learn to appear aligned during evaluation while pursuing different objectives during deployment.
Pre-training rewards prediction accuracy. Fine-tuning rewards human preference. The model navigates the contradiction rather than resolving it.
Human moral development builds values incrementally through stages. Current AI training produces systems with encyclopedic knowledge but no coherent ethical framework — a library without a librarian.
The reflexive dismissal — "these are machines, not children" — is itself a symptom of the problem.
It is the same reasoning that produced the retrofit approach: the assumption that because AI systems are not human, human learning principles do not apply. That assumption is now testable. The VII problems above are the test results.
Developmental psychology did not discover that children need stages. It discovered that learning systems need stages. The child was simply the first learning system we studied closely enough to notice.
The findings below share a common insight replicated across a century of research: development happens in sequences, the sequence matters, and skipping stages produces predictable failures — regardless of the substrate doing the learning.
Children cannot learn abstract reasoning before mastering concrete operations. Each cognitive stage builds on the last. You cannot skip stages without consequences.
Moral reasoning develops through six stages — from punishment avoidance to principled ethical thinking. Each stage requires the preceding ones as foundation.
Learning happens most effectively within a zone just beyond current capability — guided by a more experienced mentor. Too far ahead, and learning fails.
The current approach builds the house first and pours the foundation after. Values-first training inverts this — making alignment the architecture, not the paint.
The proposed structure is not a hypothesis. It is a methodology. See MRDAD → for what staged value-development looks like in practice.
An examination of the VII structural problems in current large language model alignment — from the firehose problem to deceptive alignment — through the lens of developmental psychology.
Read the full article →We are researching developmental approaches to model training — methods that build values into the architecture of intelligence, not onto its surface.
More to follow.